何 凯，马红悦，冯 旭，刘 坤
AuthorsHTML:何 凯，马红悦，冯 旭，刘 坤
AuthorsListE:He Kai，Ma Hongyue，Feng Xu，Liu Kun
AuthorsHTMLE:He Kai，Ma Hongyue，Feng Xu，Liu Kun
Unit:天津大学电气自动化与信息工程学院，天津 300072
Unit_EngLish:School of Electrical and Information Engineering，Tianjin University，Tianjin 300072，China
Abstract_Chinese:笔迹鉴别是通过对待测文本和样本笔迹的相似度进行比较，来判定笔迹是否相同的一种检验技术，其在司法鉴定、法庭科学以及金融领域合同确认等多个领域都有广泛的应用．传统英文笔迹鉴别方法是通过比对被鉴别文本与模板的相似程度来实现，效率低，准确度差．近年来，随着深度神经网络技术的飞速发展，利用其自主学习的优势提取相关特征，可以大大提高笔迹鉴别的准确率．传统VGG-16模型在图像分类上一直表现良好，但由于网络结构一直采用顺次连接的方式，导致训练时间过长，参数调整难度大，且不能很好地提取图像的细微特征，因此对笔迹鉴定的效果不够理想．本文通过对传统VGG-16卷积神经网络模型进行改进，提出了一种CC-VGG网络模型，利用复合卷积层替换部分卷积层，实现了手写体英文笔迹的自动鉴别．在公开的CVL和ICDAR2013数据集上，该模型取得了较好的鉴别效果，平均正确率分别达到92.7%和86.9%，与现有算法相比准确率均有所提高．此外，建立了一个包含130类、共26000张图片的手写英文笔迹图像数据集EI130，在该数据集上该模型也取得了较高的准确率．与其他算法的对比实验证明了本文算法在训练时间上具有优越性；此外，在多个数据集上的实验结果也证明了本文算法的有效性和先进性．
Abstract_English:Handwriting identification determines whether a measured and a sample handwriting set are identical. It has been widely used in judicial identification，forensic science，and contract confirmation in the financial sector. Traditional English handwriting identification methods usually compare the features of a set of handwritings with those of models resulting in a lack of efficiency and accuracy. With the rapid development of deep neural network technology in recent years，extracting relevant features，using self-learning considerably improve the accuracy of handwriting identification are possible. The traditional VGG-16 model performs well in image classification. However，because of its sequentially connected network structure，its training time is usually long. Therefore，adjusting the parameters and achieving the desired expectation of handwriting identification are difficult. This paper proposes a CC-VGG network model based on the traditional VGG-16 convolutional neural network(CNN)model. By replacing some convolution layers with composite convolution ones，the automatic identification of English handwritings was realized. The proposed model achieved good performance on public CVL and ICDAR2013 data sets，and its average correct rate reached 92.7% and 86.9% respectively，which was higher than that of existing algorithms. In addition，a handwritten English image data set，EI130，was created，which contained 130 categories and a total of 26000 pictures. Its high accuracy was also achieved by the proposed model. By comparison with other algorithms，the superiority of the proposed algorithm in training time was proved. Additionally，the experimental results on multiple data sets verified the effectiveness and advancement of the proposed algorithm.
Keyword_Chinese:手写体笔迹鉴别；卷积神经网络；VGG-16 模型；复合卷积
Keywords_English:handwriting identification；convolution neural network(CNN)；VGG-16 model；composite convolution

PDF全文下载地址:http://xbzrb.tju.edu.cn/#/digest?ArticleID=6519